Acute seizure risk in patients with encephalitis: development and validation of clinical prediction models from two independent prospective multicentre cohorts

Objective In patients with encephalitis, the development of acute symptomatic seizures is highly variable, but when present is associated with a worse outcome. We aimed to determine the factors associated with seizures in encephalitis and develop a clinical prediction model. Methods We analysed 203 patients from 24 English hospitals (2005–2008) (Cohort 1). Outcome measures were seizures prior to and during admission, inpatient seizures and status epilepticus. A binary logistic regression risk model was converted to a clinical score and independently validated on an additional 233 patients from 31 UK hospitals (2013–2016) (Cohort 2). Results In Cohort 1, 121 (60%) patients had a seizure including 103 (51%) with inpatient seizures. Admission Glasgow Coma Scale (GCS) ≤8/15 was predictive of subsequent inpatient seizures (OR (95% CI) 5.55 (2.10 to 14.64), p<0.001), including in those without a history of prior seizures at presentation (OR 6.57 (95% CI 1.37 to 31.5), p=0.025). A clinical model of overall seizure risk identified admission GCS along with aetiology (autoantibody-associated OR 11.99 (95% CI 2.09 to 68.86) and Herpes simplex virus 3.58 (95% CI 1.06 to 12.12)) (area under receiver operating characteristics curve (AUROC) =0.75 (95% CI 0.701 to 0.848), p<0.001). The same model was externally validated in Cohort 2 (AUROC=0.744 (95% CI 0.677 to 0.811), p<0.001). A clinical scoring system for stratifying inpatient seizure risk by decile demonstrated good discrimination using variables available on admission; age, GCS and fever (AUROC=0.716 (95% CI 0.634 to 0.798), p<0.001) and once probable aetiology established (AUROC=0.761 (95% CI 0.6840.839), p<0.001). Conclusion Age, GCS, fever and aetiology can effectively stratify acute seizure risk in patients with encephalitis. These findings can support the development of targeted interventions and aid clinical trial design for antiseizure medication prophylaxis.


INTRODUCTION
Encephalitis is inflammation of the brain parenchyma, caused by infectious or immunemediated processes, and is associated with significant morbidity and mortality despite antiviral and/or immune therapies. 1 Globally, 500 000 children and adults are affected each year. 2 The clinical presentation is variable, but typically includes acute or subacute onset of altered mental state alongside fever,

WHAT THIS STUDY ADDS
⇒ This examination of patients with encephalitis reflecting the spectrum of aetiologies from two prospective independent multicentre studies, identified that age, Glasgow Coma Scale on admission, presence of fever and aetiology were strongly associated with seizures. Using these limited parameters in a clinical prediction model, we were able to stratify seizure risk.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY
⇒ The findings can be used to support the development of targeted interventions, such as early specialist care involvement, for patients at highest risk of seizures and to aid the design of clinical trials of antiseizure medication prophylaxis.
Open access headache, new-onset focal neurological signs and, in some, seizures. 3 Seizures have particular significance as they are associated with a worse outcome and may well be amenable to prophylaxis. 4 Although seizures may be a proxy marker of severe encephalitis, there are a number of mechanisms by which they could lead to further brain inflammation and damage, including hypoxia, excitotoxicity and raised intracranial pressure. 5 However, the incidence of acute symptomatic seizures is highly variable (between 2% and 67%). 5 Although there is some limited evidence that possible risk factors include the aetiology of encephalitis, younger age and the degree of cortical involvement, our capacity to predict who is at risk of seizures remains very poor. 4 5 Consequently, there is insufficient evidence to recommend the use of antiseizure medications (ASM) as standard of care as either primary or secondary prophylaxis. 4 6 Initiation and escalation of ASM is possible in most healthcare settings, and if proven to improve outcome, could be started rapidly as ASM therapy is agnostic to eventual aetiology. Therefore, if a high-risk group were established, this could be used to stratify patients for future clinical trials of primary and secondary prophylaxis with ASM or, as a minimum, to identify which patients should be managed in settings with adequate capacity to manage this severe complication. 4 This study aims to establish the factors associated with seizures in encephalitis as well as develop and validate a seizure prediction model of clinical utility for patients presenting with an acute encephalitis syndrome, in accordance with the WHO approach. 7 8

Cohort 1 (development cohort)
Patients were recruited through the Aetiology Study of Encephalitis Study led by the UK Health Protection Agency (now UK Health Security Agency (UKHSA)) (Cohort 1). 9 The study prospectively recruited 203 patients with encephalitis from 24 hospitals in England (2005England ( -2008 serving 5 million people (11% of the English population). The study included any person of any age admitted to hospital with encephalitis, full case definition as previously published. 3 Computerised tomography (CT), MRI and electroencephalogram (EEG) were performed when clinically indicated. Clinical and postmortem samples received enhanced diagnostic testing guided by a multidisciplinary expert panel.

Cohort 2 (validation cohort)
A second cohort of 233 patients with encephalitis recruited as part of the Understanding and Improving Outcome of Encephalitis in the UK (Enceph-UK) study was used exclusively as a validation cohort for model development (Cohort 2). Enceph-UK prospectively recruited patients from 31 hospitals in England, Wales andScotland (2013-2016). Patients were eligible if they were 16 years or older and had clinically suspected encephalitis, using the same case definition as Cohort 1.

STATISTICAL ANALYSIS Seizure definition
The three outcome measures were (1) seizure occurrence at any time before or during acute admission, referred to as 'seizures', (2) the occurrence of seizures during acute admission, referred to as 'inpatient seizures' and (3) the occurrence of status epilepticus. Witness description was used to differentiate focal from generalised seizures. Further subclassification, for example, according to International League Against Epilepsy classification, was not feasible. In Cohort 2, the presence of seizures was recorded as a binary outcome under 'symptoms on admission (or in current illness, up to 8 weeks prior to admission, including prodrome)', and, therefore, description of subsequent 'inpatient' seizures was not possible.

Data extraction
Data from the first available cerebrospinal fluid (CSF) analysis were extracted. Cut-offs were taken from the UK Standards for Microbiology Investigations: Investigation of CSF. 10 Glasgow Coma Scale (GCS) was categorised as normal (15/15), mildly impaired (13-14/15), moderately impaired (9-12/15) or severely impaired (3-8/15). Outcomes were recorded according to the Glasgow Outcome Scale (GOS) 6 months after discharge from hospital. Good recovery was defined as GOS=5 and poor outcome was defined as <5, reflecting at least moderate disability. 11 Univariate All univariate analysis was conducted on Cohort 1. Categorical variables were analysed using χ 2 or Fisher's exact test. All continuous variables that were non-parametric were analysed using Mann-Whitney U test. Potential confounding variables were considered to be age, sex, ethnicity and treatment. These confounders were re-reviewed after univariate analysis and assessed for effect modification and interaction using binary logistic regression.

MODEL DEVELOPMENT
Clinical prediction modelling was designed to be applicable to routine clinical practice to stratify seizure risk in patients presenting with the clinical features of acute encephalitis syndrome. Binary logistic regression was used to ascertain predictor variables of seizures in Cohort 1 (SPSS V.26). Due to the limited proportion of patients with a clinically indicated EEG and the risk of data availability bias, EEG results were not considered for inclusion. Collinearity was assessed using correlation matrices and one of any two highly correlated variables omitted. The pattern of missing data was reviewed to assess whether data were missing completely at random. Data not missing 'completely at random' by Little's test but deemed to be missing 'at random' were imputed using multiple imputation in clinical variables with >5% missing data. Candidate variables were selected by univariate   (15) 17 (14) 14 (17)  Open access selection (p<0.25) and those identified in the literature. A selection of strongest contributing predictors was made through backward selection based on likelihood ratio. The final binary logistic regression model based on pooled estimates was converted to a provisional clinical scoring system by dividing regression coefficients of each factor by the smallest regression coefficient among the variables to the nearest integer.

Model development: inpatient seizures
To aid translation to clinical practice, a second binary logistic regression model for inpatient seizure risk was developed using the same approach to represent (1) risk at time point of admission and, (2) with maximal discrimination, and was assessed using receiver operating characteristics (ROC) curve and Hosmer-Lemeshow test (Cohort 1).

Model validation
Both scoring systems were internally validated using leaveone-out cross-validation performed in R (The R Foundation). 12 The provisional scoring system for seizures was externally validated on Cohort 2 using ROC curves, Hosmer-Lemeshow test and calibration plot.

Inpatient seizures
In Cohort 1, inpatient seizures were present in 27/43 (67%) of patients presenting with a history of seizures, and 64/124 (52%) without a history of seizures. In patients who did not present with a history of seizures, those that would go on to develop an inpatient seizure presented with a shorter duration of symptoms at 6 (1-13) versus 9 (4-21) days (p=0.034).
Reduced GCS on admission was associated with subsequent inpatient seizures. This association remained when stratifying data according to whether the patient presented with a history of seizures. In 92 patients with known GCS and without a history of seizures at presentation, those with severely impaired GCS were more likely to have at least one subsequent inpatient seizure (11/13 (85%), OR 6.57 (95% CI 1.37 to 31.5)), compared with those with moderately impaired (11/21 (52%), OR 1.07 (95% CI 0.40 to 2.93)) or mildly impaired/normal GCS (25/58 (43%), OR 0.41 (95% CI 0.17 to 0.99)) (p=0.025).

Provisional model
Presenting with a seizure and GCS were co-linear, however, GCS was most strongly associated with seizures and likely also captures whether a patient has a history of seizures due to the postictal phase, so was retained in the model. The provisional model of seizures at any time included GCS on admission and probable aetiology of encephalitis (χ2=42.53, p<0.001) (online supplemental table 2). Consistent with the univariate analysis, autoantibodyassociated (OR 11.99 (95% CI 2.09 to 68.86), p=0.017) and HSV encephalitis (3.58 (95% CI 1.06 to 12.12), p=0.096) were associated with seizures. Internal crossvalidation demonstrated 68% sensitivity, 72% specificity, with a positive predictive value (PPV) of 62% and negative predictive value (NPV) of 77%, with overall accuracy 70%.
Inpatient seizure risk: SEIZURE score A second binary logistic regression model was developed to identify predictors of inpatient seizures based on the information available on admission and then on these parameters combined with aetiology once established (Cohort 1) (table 2). The derived, points-based SEIZUre Risk in Encephalitis (SEIZURE) score stratified risk by decile and is designed to be applied by healthcare professionals when a patient of any age with suspected encephalitis is admitted to hospital with two weighted scoring systems for application prior to and following identification of the probable aetiology ( figure 4). Internal crossvalidation demonstrated 66% sensitivity, 72% specificity, PPV 69% and NPV 69%, with overall accuracy 69%. Patients in the highest risk category on admission had an OR 7.17 (95% CI 2.55 to 20.16) of seizures compared with those in the lowest risk categories and an OR of 15.51 (95% CI 5.60 to 42.96) once probable aetiology was established (table 3).

DISCUSSION
Acute seizures affect many patients with encephalitis and are associated with increased need for intensive

Open access
care support and a worse outcome, and moreover may further contribute to brain injury through excitotoxicity, host immune responses and raised intracranial pressure. 5 12-15 However, there are currently no established tools to stratify patients as to their risk of seizures and status epilepticus. Without such risk stratification, it is currently not possible to identify which patients would be best managed in centres with adequate clinical facilities, such as those with intensive therapy units and continuous EEG monitoring, and also to identify whom might benefit from primary and secondary ASM prophylaxis.
Through our evaluation of two independent prospective multicentre cohort studies, we identified multiple factors associated with increased risk of seizures during the acute illness, particularly low GCS on admission, fever and an autoantibody-associated or HSV aetiology. Patients with seizures presented to hospital earlier, even in those who have not yet had their first seizure at the time of presentation. Models for seizure risk could be established which, despite requiring a small number of variables, were strongly predictive of acute seizures. Low GCS on admission was more strongly associated with inpatient seizures than whether the patient presented with a seizure history or not. The provisional score accurately determined seizure risk in the first cohort, but potentially overestimated seizure risk in the second cohort, perhaps reflecting under-documentation of seizures in this cohort as seizures were limited to those documented at presentation. The 'SEIZURE score' for inpatient seizures requires further external validation. Improved access to easy Open access EEG-monitoring on a wider scale, or the establishment of novel biomarkers could enhance the accuracy of riskstratification. In addition, the impact of ASM prescription as primary and secondary prophylaxis requires further evaluation. 4 5 There are likely to be multiple structural and biochemical mechanisms underlying seizure risk in patients with acute encephalitis syndrome. For example, HSV encephalitis has a predisposition to affect epileptogenic areas within the frontotemporal region and in autoimmune encephalitis, the antibodies associated with neuronal cell-surface antigens, that are highly expressed in this region, are themselves often directly involved in the disease process. [16][17][18] The differential disease mechanisms observed in specific aetiologies of encephalitis may influence seizure risk, however, there was inadequate power in this analysis to establish factors associated with seizures within aetiological subgroups. Certain clinical features were less common in patients with seizures, specifically, stiff neck, photophobia, lethargy and any focal deficit on neurological examination. This may in part be explained by aetiological distinctions, as focal neurological deficits were most frequently reported in encephalitis caused by Varicella zoster virus and ADEM, which were least strongly associated with seizures; and particularly the latter which is associated with subcortical white matter lesions as opposed to cortical inflammation which drives seizures. 19 Many clinical features associated with seizures in our analysis are likely to be proxies for underlying mechanisms, rather than risk factors in themselves. Nevertheless, these features can inform future mechanistic studies, particularly through in vivo models of encephalitis. 20 EEG abnormalities were strongly associated with clinical seizure activity and EEG also identified three cases of NCSE, which is increasingly recognised in encephalitis, particularly autoimmune aetiologies. 21 22 Status epilepticus in the context of encephalitis is frequently refractory and has a poor outcome. 23 24 Patients with seizures were less likely to have CSF pleocytosis or low CSF glucose. Given that lumbar puncture is contraindicated until patients are stabilised following a seizure, we hypothesise that this result could be related in part to delayed lumbar puncture, or also may reflect the increased proportion with autoimmune encephalitis in this group. 19 These CSF parameters likely reflect aetiological distinctions rather than direct biomarkers of seizures. Nevertheless, these data sets did not provide sufficient granularity to determine the CSF white cell count relative to the time from/before a seizure and this requires further study. No relationship was observed between the presence of normal or abnormal CT or MRI findings and seizure risk. It may be that the imaging variables in this analysis were too crude as they were based on retrospective interpretation of clinically indicated scans performed at multiple sites. Volumetric analysis for research purposes of specific brain regions or structures would be more sensitive. 6 25-27 The finding may additionally reflect the high incidence of seizures in those with autoimmune encephalitis, who often have normal or near-normal neuroimaging. A single-centre study of 94 patients in China found cortical or hippocampal abnormalities on neuroimaging independently predicted progression to super-refractory status epilepticus. 28 Notably, the potential associations of seizures identified in our study; aetiology, GCS and younger age, were also reported in a single-centre study in Northern India, despite large differences in the cohort. 26 The likelihood of a seizure being witnessed may confound associations, for example, younger children in community settings may be more likely to have a seizure witnessed.

Open access
Our study corroborates previous work demonstrating an association between seizure activity and poor outcome in encephalitis. 14 29-31 Although seizures may be a proxy marker of severe disease, there are a number of mechanisms through which seizures could cause further brain damage. Seizures cause significant systemic metabolic and biochemical disturbance including hypoxia, hypoglycaemia, metabolic acidosis as well as direct central nervous system perturbations including glutaminergic activity, raised intracranial pressure and blood-brain barrier permeabilisation, as well as low CSF glucose and high CSF lactate. 32 A study of 144 patients with Japanese encephalitis presenting to hospital in Vietnam, showed that patients with recent seizures had high CSF lactate:glucose ratios and high CSF opening pressures and that patients with opening pressure >25 cm were more likely to die. 14 A more recent analysis of CSF biomarkers in HSV encephalitis indicated that acute inflammation may drive subsequent synaptic autoimmunity and proposed a trial of post-acute corticosteroids. 33 In addition, several inflammatory markers have been associated with a lower GCS, increased oedema and a worse outcome in encephalitis, especially the interleukin-1 family relative to their endogenous antagonists. 13 It remains unknown whether seizures intervene with the underlying encephalitic process.
Despite the high prevalence and prognostic importance of seizures, the most recent Cochrane review concluded that there is insufficient evidence to support or refute the routine use of antiepileptic drugs for the primary or secondary prevention of seizures in viral encephalitis. 4 A recent randomised controlled trial of secondary prophylaxis for acute symptomatic seizures in children with encephalitis demonstrated that a 4-week course of ASM was comparable to 12 weeks in terms of the incidence of seizure recurrence. 34 A rabbit model of HSV-1 encephalitis, untreated with aciclovir, showed that all animals that had a seizure became moribund or died, but that phenobarbital prevented seizures and significantly reduced mortality. 35 Further questions remain regarding choice and duration of antiepileptic agents. 6 Any intervention strategy would need to consider the high baseline risk of seizures in patients with encephalitis and the presence of subtle and subclinical seizures including NCSE. 14 21 36 Our findings reflect two relatively large prospectively recruited cohorts but have limitations, principally due to the retrospective nature of seizure-focused analysis. The observational nature of the study has an intrinsic risk of confounding which we have attempted to address in both the analysis and interpretation of results but there may be a residual impact. Additionally, the data were collected from 2005 to 2016 and seizures were not the primary focus of the initial data collection, potentially resulting in missing data. The diagnosis and management of encephalitis may have changed over this time period, particularly the identification of autoantibodies. These factors are balanced by the substantial sample size for a relatively uncommon condition, the granularity of the UKHSA data and the enhanced diagnostic testing performed.

CONCLUSION
These finding indicate that patients with seizures during encephalitis present earlier, but despite this, have a worse outcome, suggesting there may be a window of opportunity for intervention that is currently not being exploited. This study provides a foundation for risk stratification of seizures in encephalitis on clinical grounds alone. Biomarkers and improved access to EEG-monitoring could enhance model accuracy and allow for the development of targeted interventions. The SEIZURE score can be used to aid the design of clinical trials of primary and secondary prophylaxis with ASM. Contributors JG, TS and DWGB initiated the original HPA study, obtained funding and recruited hospitals to the study. JG led HPA study design. GKW and BDM designed secondary analysis. GKW did the literature search. GKW, BDM, KT and GB did the statistical analysis. GKW and BDM developed the figures. GKW, TS and BDM wrote the first draft. BDM is the guarantor of the paper. All authors contributed to interpretation of data and critically approved the final paper. Disclaimer The sponsors of the study had no role in study design, data collection, data analysis, data interpretation or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.